HIGH PRECISION DEM GENERATION FROM SPOT STEREO IMAGERY BY OBJECT SPACE LEAST SQUARES MATCHING K. C. LO , N. J. MULDER LT.C. P.O. Box 6, 7500 AA Enschede The Netherlands (ISPRS Commission III) Abstract: For automatic DEM generation from multiple view SPOT imagery with high precision, an approach of "Refinement from Coarse" is proposed. In the coarse DEM generation stage, in contrast with the traditional methods which match at intensity level in image space with the signal processing technique, the concept of Knowledge Engineering is used to perform high level Feature Matching by Property List and String Matching for Correspondence analysis with high reliability. In the refinement stage, the Object Space Least Squares Matching method, which is the most rigorous method from a theoretical viewpoint, is proposed for coarse OEM refinement. This method is different from the traditional two-step approach which matches the corresponding point in image space first, and then determines the DEM by Space Intersection; this method is improved by back mapping the image data into object space to get object reflectance O(x,y) with referring the object surface Z(x,y), and perform matching in Object space. It is simultaneously to determine two functions in the object space: the terrain relief Z(x,y) and the terrain reflectance O(x,y) in one solution with least squares adjustment iteratively. The disadvantages of matching in image space which the multi view image of the same object has different geometric or reflectance distortion can be avoided. It's flexibility allows the user to handle more than two SPOT multi-view images in one solution, which increases accuracy and reliability as well. It is good for SPOT images which offer the resolution with 10 meters groundel size only. Key Words: OEM, Correspondence Analysis, Property List, String Matching, Object Space Least Squares Matching. I. INTRODUCTION automatic generation of coarse DEM data by Linear Feature Matching. In order to plan and execute complicated sequence of operations and functions, we believe the methods of Knowledge Engineering should be used, and correspondence Analysis will be based on object detection with geometric definition and object description by means of Property Lists [Mulder et al.,1988]. The selection / representation and use of proper knowledge is a central problem in research. A range of different knowledge representation techniques must be developed, along with a number of approaches to applying knowledge, which are concerned in .the field of Meta-Level Knowledge. The generation of a GIS for a 3-D object oriented data base is required urgently in many countries. The automatic extraction of 2.5-D information from SPOT stereo images is, potentially, an efficient and economic way. Some on-line (real time) commercial systems have appeared in prototype, but the methods of the off-line system for reconstructing the earth surface with high quality and acceptable economy must be researched and further developed still; We want to establish a system for generating DEM with high accuracy / high quality / high reliability with the support of the method of Object Space Least Squares Matching. It will then offer good fundamental 2.5-D information to GIS for mUlti-purpose applications. 2. 2.1 2.3 Traditional matching in Image Space has the shortcoming that the same object in multi-view images appears with different geometric distortions and radiometry. The geometric distortions are caused by the central perspective which produces relief displacement in the image; the tilt of the sensor (SPOT has off nadir angle from a to 27 degrees) causes tilt displacement in the image. As the multi-view SPOT images are taken at different positions, at different times or under different illumination conditions, this produces different radiometry for the same object in different images. These geometric and radiometric differences produce matching failures or reduce the accuracy. Therefore, motivation for improved developments should come from the realization that all information in images is inherent in the object space, and the transformation of the matching problem from Image Space to Object Space leads to a unified and precise approach where all available knowledge is referenced to the same basis. Research has to be carried out into ways of how to model the surface of the terrain and its reflection which is suitable for matching properly and efficiently. On the other hand, high accuracy can be obtained by using Minimum Cost Matching; e.g. the Minimum Euclidean Distance can be selected as the "Cost" in Euclidean space for similarity assessment. Because, in case the distance is small, the distance behaves as SIN function of the angle which is between the Feature Vectors for matching, it BACKGROUND PROBLEMS SPOT Satellite Imagery Since the SPOT imagery is acquired by push-broom scanners, the imaging geometry is different from the conventional central perspective photographs with frame camera, and the orbit parameters require simulate dynamic modelling. For solving the inverse camera model problems, the point in issue is how to determine the orientation parameters of each scan line with sufficient accuracy (e.g. improving the accuracy of Tie Point measurement/transfer and best pattern of Ground Control Point distribution) and at lowest cost (e.g. minimum number of Ground Control Point's). On the other hand, the CCT of SPOT not only presents imagery data, but also ot'fers special on board auxiliary data, such as information about position, attitude, look direction, radiometric calibration of scene/sensor [SPOT User's Handbook, 1988]; how to fully use this information for obtaining benefits in Aerial Triangulation ( A.T.) and DEM generation stage must be considered in every phase of image data processing. 2.2 Object Space Minimum Cost Matching Knowledge Engineering The existing methods of signal correlation and feature matching are limited for handling some special correspondence analysis problems, such as 133 may offer a sensitive measure for the distinction of extreme similarity in the primitive feature space. For the Least Squares Matching method, although the approach is completely different (i.e. not searching / comparing and selecting, but simultaneous Solution / determination of the unknowns), the basic idea is similar, namely, matching with very high accuracy to the degree of subpixel. This also has the advantage that Least Squares Adjustment is flexible which allows the users to perform matching with more than two images which can increase accuracy and reliability, and offers the theoretical quality estimation of the result as well. 3. 3.1 technique, in order to solve the problem of high correlation between orientation parameters caused by narrow FOV of SPOT (4.125 degrees only) [Chen & Lee,1989] [Shibasaki, el at.,1988]. At the same time, we try to reduce the number of unknown orbit parameters (e.g. simulating the orbital model for position by 2nd degree polynomial, and linear polynomial for attitude) resulting in a reduced number of GCP, and aim at finding out the best distribution of position of GCP in the adjustment of A.T. providing sufficient accuracy for later matching. d) Application of Object space Least Squares Matching with exterior orientation parameters as unknown; using on board auxiliary data as initial value, try to perform highly accurate Tie Point Transfer with interaction in adjustment of A.T. by iteration. This is the most difficult part to solve, because the known exterior parameters are the back bone of Object Space Least Squares Matching. If we treat them as unknown with on board data as initial values instead, we need to know how good the initial value should be to make the iteration convergent. THE SYSTEM COMPONENTS OF THE APPROACH Preprocessing 3.1.1 Coordinate System Transformation a) For mapping purposes, the coordinates of Ground Control Point (GCP) are offered in the Mapping System, because a SPOT image covers an area of 60 km x 60 km, hence the effect of the earth curvature can not be neglected any more. Therefore, a transformation from the Map Coordinate System to the Geographic System, and further to the Geocentric System is necessary. However, as the number of digits of coordinates in the Geocentric System is large, and.a Double Precision is required in the data process to avoid truncation error, it is necessary to transfer the coordinates to Local Tangent Plane System to reduce the number of digits, and thus to save memory and speed up processing. e) Empirical accuracy study of exterior orientation parameters from A.T. with the application of previously mentioned techniques, trying to get sufficient accuracy to meet the requirement of Object Space Least Squares Matching. 3.2 Coarse DEM Generation by Correspondence Analysis with Property List a) There are several methods to enhance the Linear Features ( Line / Edge ) and then extract them; however I the original position and intensity of linear features should not be changed if the features will be used for matching (not for visual satisfaction) later. Therefore, a non-linear filter, such as Conditional Rankorder Filter [Mulder & Sijmons, 1984], can be selected for segmentation; thereby the enhancement of features is done by smoothing the background (suppressing the minor features / noise also) and keeping distinguished Linear Features· in their original situation. b) For using the on board data which are in the Geographic System / the Geocentric System / the Local Orbit Reference System / the Local Attitude Reference System, transformation to the Local Tangent Plane System is needed also. 3.1.2 Region Matching for DEM Generation and Change Detection Because the matching would fail within the homogeneous intensity region, or in the regions which the land cover has been changed when the satellite stereo pairs are taken with a long interval of time. Therefor, we use Conditional Rankorder Operator to smooth the intensity within the region first, then start Region Growing for image segmentation, the boundary of region can be extracted, and the shape can be described by V-s Curve, combine with other properties of region, such as the area, the position of gravity centre, etc., to form a Property List. The initial probability of region matching can be obtained by minimum cost function with the weighting properties in the list. Then the matching probability are being adjusted by Relaxation Processes. until the final conjugated region pairs are determined. The elevation of the region can be calculated with the conjugated region, and the change detection can be done by checking the mismatching regions and comparing the intensity between the conjugated region after region matching, the matching failure problem can be solved in these area [Lo,1992]. b) Reduce the 2-D search to a 1-D search during matching, by the resampling of image data into parallel line pairs (the approximate Epipolar Line pairs). This is, however, more difficult to apply to SPOT images, because the orientation parameters are a function of time [Otto,1988] [Zhang & Zhou, 1989] • c) Apply a Gradient Filter to the image and detect the Linear Features with Zero-Crossing. d) Property List Formation by collecting the properties of Linear Features such as Position, Amplitude and Shape of peak / valley in intensity profile along the parallel line pairs [Lo,1989], and the Orientation of Linear Features [Kostwinder el at, 1988] • 3.1.3 Aerial Triangulation ( SPOT Orbit Determination) a) On board data are used to define the overlap area of multi-view images and to choose the Tie Points at the proper position and evaluate / select the specific image properties (e.g. high contrast in X and Y direction) Le. the most suitable features for matching required for automatic point transfer. After image segmentation is performed, the crossing of lines / edges are detected as GCP, and sufficient properties (structure representation) of GCP can be obtained for GCP identification by the Line-Based Structure Matching Method or Chain-Coding Matching. Moreover, on board auxiliary data and ground coordinates of GCP are also used to help automatic identification of GCP with property list for the correspondence analysis between maps/photos. e) To offer the criterion for Correspondence Analysis, Cost Function Modelling is required by assigning different weights to the individual properties according to it's reliability and major/minor contribution to express the characteristic of feature. The weight can be assigned by prior analysis or by experiment with Trial and Error. f) Between the conjugated parallel line pairs, the corresponding linear feature can be extracted by String Matching which selects the Minimum Cost as best matching, based on information from the Cost Function. For the conventional matching strategy, the Target Area of the left image is selected to search for the best match in Search Area of the right image only; the result may be different, however, if the matching is from right to left. The String Matching uses the mutually matching strategy which matches not only left to right but also right to left, then selects the real Minimum Cost among them as best matching with the marking technique for extracting them. It increases the reliability of the result. If we confirm the extracted linear features again by checking the b) Establish the model for simulating the short arc orbit of SPOT as a function of time [Konecny et al. ,1987) [Kratky,1988). c) Extract on board data from CCT of SPOT [SPOT User's Handbook, 1988) and use them as constraint (e.g. the attitude data) by the Pseudo Observation 134 continuation of linear features between neighbouring parallel lines, the reliability can be increased still more [Lo, 1989] . The extracted corresponding linear features pairs can be used to generate coarse DEM, the major requirement of generating coarse DEM with high reliability is fulfilled. 3.3 iteratively [Wrobel,1988] [Helava,1988] [Ebner & Heipke,1988] [Heipke,1990]. By this way, the digital image matching, DEM generation and orthophoto computation have been combined into one approach. The main principle of Object Space Least Squares Matching is as follows: The basic Observation Equation of every pixel within the patch for matching by the Least Squares Adjustment is: Refinement of Coarse DEM Data by Object Space Least Squares Matching L= R D* v 3.3.1 The Illumination and Reflection Model for facet stereo matching There are three major phenomena of the reflection : Ambient reflection, Diffuse (Lambertian) reflection and Specular (Mirror) reflection. It describes the relationship of reflected radiance of a small facet of the surface, the specific viewing direction, the complex illumination of the scene and light reflecting properties of the material. (Weisensee,1988) proposed three models of light reflection which can be used for Object Space Least Squares Matching, they are derived from general model as follows: L R) cose - D( Z, A, (8) f v noise or residual error in intensity the unknown intensity value assigned to one groundel e the angle between the surface normal and the direction of sun D( ): back mapping the image intensity value of corresponding pixel to the groundel. Z the heights of N x M grid points (DEM) to represent the terrain surface Z(X,Y) which can offer the height in any position by bilinear interpolation. A orientation parameters of the sensor L illumination model R object reflectance model a) The influence of L & ~ can be simplified ey local linear radiometric correction with offset 81 and gain 52 (But not for close range photogrammetry and large scale photo). Therefore, (8) can be: D (1) the reflected radiance cause by incident radiance of illumination the ambient component from half space L 1a and the semi-spatial reflectance Ra n light sources having irradiance LIn the surface normal the direction to light source n the apparent solid angle of incident radiance from a particular light source the diffuse reflectance Rd and its fraction of proportion for diffuse reflectance the specular reflectance R and its fraction of proportion for spe~ular reflectance v D* cose - 51 - 52 * (9) D(Z,A) For the patch of first image, 81 and 82 can be assumed as constant I they are unknowns for the patch of second, third and further images. b) For Point Transfer in A.T., Z and A are unknowns also, expanding the nonlinear part (82 * D(Z,A» of (9) !:o li~ear increments from approximate values: ( S1 , S2 ) 0 = ( 0 ,1 ), we obtain: In our case, we can assume that there is only one light source (n=l,the sun) to be considered. The amount of light reflection which independently with the viewing direction is : v The the the and (2) the specular Component which is rare exist and can be omitted is: (3) the relationship between recorded intensity G; of discrete pixel i by the sensor and the reflected radiance from the terrain surface can be a linear transformation: = D* cose - 81 - 82 * D(Z,A) - 82 * [(aD/aZ)o dZ -(aD/8A)0 dA] _oD(Z,A~o dS2 (10) coefficients indexed by 0 are calculated with estimated values which will be updated during iteration; Expanding the coefficient (aD/aZ)o (aD/aA)o of the design matrix: (aD/aZ)o (aD/ax .ax/aZ)o+ (aD/ay .ay/az)o (11) (aD/aA)o (aD/ax .aX/aA)o+ (aD/ay .qy/aA)o (12) Because aD/ax and aD/ay represent the gradient of the intensity, it can be used for a quality estimation (selection), and provide the different weights accordingly in Least Squares Adjustment. (4) c) In case of DEM generation after A.T •• The orientation parameters of sensor A are known already. Therefore, the observation equation for each pixel within the matching pat~h is: if the terrain surface is perfect ambient reflection, the first reflection model can be: (5) v D. is the reflected radiance of facet i, if the t~rrain surface is assumed as perfect Lambertian reflectance, the second model is: G;= a j + b(cos(NQ.lDQ D; = a j + bfcose;Q D; Q D; cosO - 81 - 82 * D(Z,A)o - 820 * (13) (aD/aZ)o dZ - D(z,R)o Ds2 !f we want to obtain DEM Z(X,Y) only, the unknown D can be eliminated (Reduced Normal Equation) during the inversion of the normal equation to save calculation time. (6) the third model: G; = a j + b k = D* 3.3.3 The characteristics of Object Space Least Squares Matching a) We perform the matching in Object Space instead of in Image Space r because the disadvantage. of matching in image space would be that the multi view image of the same object would have different intensity or shape caused by different relief displacement / tilt displacement / different illumination situation / different reflection effects from different positions of the sensor relating to different relief of object etc., resulting in matching failure or poor accuracy. (7) results from applying the transformation (4) to facets instead of windows. It means a; is still valid for window jf the b k applies to a bilinear height/reflection facet k respectively. 3.3.2 The orinciple of Object Space Least Sguares Matching Back mapping of image data into object space to get object reflectance D(x,y) (image inversion) by referring the object surface Z(x,y) which is approximate at the beginning, and perform matching on the object surface. It simultaneously determines two functions in the object space: the terrain relief Z(x,y) and the terrain reflectance D(x,y) with Least Squares Adjustment b) High accuracy can be achieved by using the Minimum Cost Matching which is based on the Theory of Minimum Cost Sequence of Error Transformation. This method may select the Minimum Euclidean 135 Distance as the "Cost" in Euclidean Space for similarity assessment. Because the distances behave as SIN function of angle between the feature vectors in case the distance is small, it offers a sensitive measure for the distinction of close similarity in the primitive feature space. In the Least Squares Matching Method, though the approach of matching is totally different, the basic idea is similar (Min. ~vv <--> Min. distance), namely matching with very high accuracy to the degree of subpixel; it requires, however, a very good conjugacy prediction (coarse DEM). The commonly used Normalized Cross Correlation method provides a correlation function behaviour as a COS function [LO,1991]. This function has a flat peak at maximum similarity; therefore, to interpolate this function to get subpixel accuracy is hardly worth the effort. The other way is called Indirected Pixel Transformation: we start with a coarse DEM grid (X,Y,Z), and the Collinearity Equation with orientation parameters of the scanner is used to get the corresponding pixel position (x,y). Another problem arising from SPOT's Push Broom scanner is that the orientation parameters of each scan line are different (for a frame camera, it is the same in the whole image); therefore, if we don't know which set of orientation parameters we should use for transformation, we can not use the Collinearity Equation to get the corresponding pixel on the image and transfer its intensity (after resampling) to that grid point. If we can solve this problem, the Indirected Pixel Transformation is better than Directed Pixel Transformation, as the weighted average of pixel intensities which are obta~ned from multi-view images according to the same grid point, can be used as the estimated object reflection D (X, Y), and the weight is assigned according to the slope of ray. c) When we use the unknown parameters to establish a mathematical model (Observation Equations) which can precisely describe the phenomena of observation/ sampling (e.g. the back mapping of intensity from image space to object space), the Least Squares Method minimizes the differences between the description model with actual model to determine the value of unknowns ( when the observation equation is linear ), or to update the estimated values of unknowns iteratively (when the observation equations are non-linear). Every pixel involved in matching can provide one observation equation; if redundant observations exist (the number of observation equations is more than the unknown parameters, i.e. the over-determined problem), the Least Squares Method is a good method to determine the value of unknown parameters even when the errors of observations are not a Normal Distribution. The principle of the Maximum Likelihood Method is identical with the principle of Least Squares Adjustment, if we have assumed that the errors of observations are a Normal Distribution. But unlike the Maximum Likelihood Method of estimation, the method of Least Squares does not require the knowledge of distribution from which the observations are drawn for the purpose of parameter estimation; however, for testing of hypotheses, we would require the knowledge of the distribution [Bouloucos,1989]. f) The Least Squares Matching method is capable to handle any number of images over two, e. g. the Triplet, as well as images scanned in various spectral bands simultaneously. It increases both reliability and accuracy of the result [Shibasaki & Murai,1988]. g) The traditional method uses windows of pixels for matching to determine a single point (usually the middle point) only, but Obj ect Space Least Squares Matching uses a window of pixels for matching to determine multi-points in a grid pattern DEM in one solution. If preprocessing provides prior knowledge about the quality of matching windows (e.g. the gradient of intensity), we can assign different weights accordingly (e.g. give the high contrast pixels a larger weight) in the Least Squares Adjustment. This means that we require the high contrast pixels to offer a larger contribution for the decision making which helps to avoid making the wrong decision in the homogeneous part of the image. Thus, a combination of advantages from Feature Based Matching and Area Based Matching can be obtained. DEM determination executed in this way is thus called Multi-Point Matching, and offer higher reliability [Rosenholm , 1987b]. d) Least Squares Window Matching has been used in the two-step approach whereby matching in image space is performed to get the conjugated position first, then the Space Intersection is used to reconstruct the object surface [Ackermann, 1984 ] . Its window size is limited to 20x20 pixels or 30x30 pixels; if the window is smaller than this, reliability is decreased, but if it is larger than this, accuracy will be poor because matching is performed on image space, and the geometry model is simplified [Rosenholm,1987a]. The improved approach unifies these two steps into one, and matching on object space. Back mapping of image intensity into object space to get object reflection D(X,Y) (image inversion) is done by referring the coarse object surface Z (X, Y), then perform matching on object surface and refine the coarse object surface iteratively. In addition to this, two functions in the object space, i.e. the object surface Z(X,Y) and the object reflectance D(X,Y) are simultaneously determined (considered) in one solution with Least Squares Adjustment; this is the reason why we consider this to be a more rigorous method than the earlier methods. h) Robust Estimation techniques can be applied in Least Squares Adjustment to get rid of noise as a gross error of sampling (observation). i) The Least Squares Method provides theoretical quality estimation of the matching result based on statistics theory. It also offers useful information for cleaning gross errors in DTM data in the postprocessing stage as quality control, as well as for using the DTM data in GIS [Day & Muller ,1988] . j) The grid DEM is separately generated patch with the quality estimation. The tion of the whole DEM can be done, e.g. of the Finite Element Method, with a improvement of DEM in the last stage. al., 1988]. patch by aggregaby means quality [Xiao et 4. SUMMARY 4.1 Summary of the Matching Selection of Approach Algorithms and the There are many matching algorithms that can be used; we summarize the relevant algorithms, and indicate those selected and applied in this system with the mark of ***. e) By referring the coarse object surface Z(X,Y), there are two ways to back mapping of the image intensity into the object surface in order to get the estimated object reflection D(X,Y). One is called Directed Pixel Transformation: it starts with pixel position (x,y) in the image; the Collinearity Equation with orientation parameters of its scanner is used to intersect the coarse object surface Z(X,Y) to get the position of corresponding groundel X,Y, and transfer the intensity of this pixel to it. However, the problem is that the groundels distribution on object surface after back mapping present a random pattern and they are different in different multiview images also; another shortcoming is that the height of these random position points needs to be interpolated to grid DEM but lose information. (a) Information for Matching: *** Intensity-Based *** Feature-Based (Property-Based) (b) Criterion for Similarity Assessment: * Angle between Matching Vectors: COS Function ---> Less Accuracy/ High Reliability *** Distance between Matching Vectors: SIN Function(in small distance) ---> High Accuracy / Low Reliability 136 (c) Matching strategy: stereo compilation of mapping can intervene. Therefore, if we can establish a Knowledge Base in which there is knowledge such as the well trained operator has, and incorporate it into the off-line system as an Expert system, we may obtain a similar capability as the on-line system with an operator. Hereby, we need to implement our approach to the computer in a semi-automatic / interactive mode with human interference in order to gain more experience and enough knowledge to enter into a Knowledge Base; then the problems of automatic DEM generation can be solved by an Expert System with sufficient knowledge in its Knowledge Base. Some primary experiments have been done for system start up/ interior orientation/ relative orientation in the measurement stage to perform diagnostics during the process [Kretsch, 1988]; there is still a long way to go for handling the whole system by a mature Expert System. Therefore, we have to start implementing our approach in the computer to accumulate our knowledge and improve the Expert System. * One way Search / Compare and Select from Target Window to Search Window *** Two ways (mutually) Search / Compare and Select between each other *** Solve / Determine the Unknowns * Multi-Pixel Matching ---> one point determination *** Multi-Pixel Matching---> Multi-points determination ( Weighting the good / high contrast pixels to help the poor pixels result in higher Reliability ) (d) The Minimum Cost Sequence of Error Transformation: * City Block Distance (Min. Absolute Difference) *** The Min. Euclidean Distance in Euclidean Space(Isotropic) (The Weighted Min. Distance in Feature Space) . *** The Least Squares Adjustment(~vv --> M~n.) with unequal weight REFERENCES summary of the process stages and their purposes 4.2 [ 1] preprocessing .Coordinate System Transformation from Map System to Local Tangent Plane System *Region Matching/OEM Generation oja-- Board Data S [2] Tie Point Selection/ GCP Identification (Map/Photo) [2a].Assessment for Good Image [2b] . Structure Feature Detection Structure Feature Representation Structure Feature Matching [1] Ackermann, F. [1984]: Digital Image Correlation Performance and Potential Application in Photogrammetry. Photogrammetric record, II (64): pp. 429-439. [2] Bouloucos, T. [1989]: Applied Statistics and Quality Control. ITC Lecture Note. pp. 60-63. [3] Chen, L.C./ Lee, L.H. [1989]: Least Squares Predicting using On-Board Data in Bundle Adjustment for SPOT Imagery. 12th Canadian Symposium on Remote Sensing. Vancouver, Canada.Vol.2, pp. 450453. [4] Day, T./ Muller, J.P. [1988]: Quality Assessment of Digital Elevation Models produced by Automatic Stereo Matchers from SPOT Image Pairs. ISPRS 16th Congress, Kyoto, Com. III. pp. 148-159. [5] Ebner, H./ Heipke, C. [1988]: Integration of Digital Image Matching and Object Surface Reconstruction. ISPRS 16th Congress, Kyoto, Com. III. pp. 534-545. [6] Heipke, C. [1990]: Multiple Image Matching in Object Space. International Archives of Photogrammetry and Remote Sensing. Vol.28, Part 3/2, pp. 294-302. WuHan. [7] Helava, U.V. [1988]: Object Space Least Squares Correlation. ISPRS 16th Congress, Kyoto, Com. II. pp. 321-331. [8] Konecny, G./ Kruck, E./ Lohmann, P./ Engel, H. [1987]: Evaluation of SPOT Imagery on Analytical Photogrammetric Instruments. PE&RS Vol. 53, No.9, pp. 1223-1230. [9] Kostwinder, H.R./ Mulder, N.J./ Radwan, M.M. [1988] : DEM Generation from SPOT Multiple View Images using Correspondence Analysis. ISPRS 16th Congress, Kyoto. Com.III, pp. 111.442-449 (B11). [10] Kratky, V. [1988]: Universal Photogrammetric Approach to Geometric Proceeding of SPOT Images. ISPRS 16th Congress, Kyoto, Com. IV, pp. 180-189 . [11] Kretsch, J.L. [1988]: Applications of Artificial Intelligence to Digital Photogrammetry. Purdue University, Ph.D Thesis. [12] Lo, K.C. [1989]: Image Quality Assessment for Automatic Production of DEM Data. Conference of Photogrammetry Automation, Tainan, Taiwan. pp. J1-J20. [13] Lo, K.C. [1991]: Review and Analysis of Stereo Matching. Journal of Surveying Engineering. Vol.33, No.3. pp. 23-42. [14] Lo, K.C. [1992]: The Automatic DEM Generation and Change Detection by Region Matching. ISPRS 17th Congress, Washington,D.C., Com.III. [15] Mulder, N.J./ Sijmons, K. [1984]: Image Segmentation using Conditional Rankorder Filtering. ISPRS 15th Congress, Rio de Janeiro. Com. III. [16] Mulder, N.J./ Radwan, M.M. [1988]: Correspondence Analysis applied to Image Matching and Map Updating. ISPRS 16th Congress, Kyoto, Com. IV, pp. IV.422-429 (B11) [17] otto, G.P. [1988]: Rectification of SPOT Data for Stereo Image Matching. ISPRS 16th Congress, Kyoto, Com. III. pp. 635-645. [18] Rosenholm, D. [1987a]: Empirical Investigation of Optimal Window Size using Least Squares Image Matching. Photogrammetria (PRS),42, pp. 113125. *Avoid: Earth curvature .. Double Precision *Avoid Matching Failure in homogeneous area *Apply Knowledge Engineering to increase the degree of Automation *Offer Orbit Parameters with Min. No. of GCP and best ~----------------------~ distribution of [3a] . Bundle Adjustment with OnGCP. '----/-----'1 Board Data as the Constrain * Improve the [3b] . using Object Space Least Accuracy of Squares Match with Orientation Point Transfer Parameter of Scanner as unknown then improve the to perform the highest accuracy Accuracy of A. T. Point Transfer interactively also. with A.T. by iteration rbit Data [3] Aerial Triangulation / Tie Point Transfer [4] f------+- Coarse OEM Generation by Feature Matching *. Correspondence Analysis: a. Image Smoothing b. Image Rectification c. Linear Feature Detection d. Property List Formation e. String Matching for Linear Feature Extraction * . Space Intersection for OEM Determination *Automatic Generation of Coarse OEM Data with High Reliability [5]Refinement of Coarse OEM Datal *High quality 'I OEM Generat~on • . Object Space Least Square with Matching subpixel Accuracy 5. CONCLUDING REMARKS In general, computer stereo vision belongs to the class of Ill-Posed inverse problem. Although the Obj ect Space Least Squares Matching is the most rigorous method with high precision from a theoretical viewpoint, but it is a serious Ill-Posed Problem, and is difficult to implement in practice, therefore, every method to improve the computational stability of image matching is very important, e.g. try to include all available geometrical constrains, such as the regularization method to minimize the surface curvature is tried to help the convergency of solution; the idea of pyramidal approach is used to improve the problem that the range of convergency is very small, etc .. On-Line system for automatic DEM generation can be used in the map production line because an operator is still involved. It means that all the matching algorithms still can not completely solve all problems because the images of terrain and the terrain itself are so complicated. However, when matching has failed, a well trained operator who has sufficient knowledge about problem solving of 137 [19] Rosenholm, D. [1987b]: Multi-Point Matching Using the Least Squares Technique for Evaluation of Three-Dimensional Models. PE&RS, Vol. 53, No.6, pp. 621-626 [20] Shibasaki, R./ Murai, S. [1988]: Improvement of Mapping Accuracy by Applying Triplet Matching to SPOT Imagery. ISPRS 16th Congress, Kyoto, Com. III. pp. III.264-273. (B10) [21] Shibasaki, R./ Murai, s./ Okuda, T. [1988]: SPOT Imagery Orientation with Satellite Position and Attitude Data. ISPRS 16th Congress, Kyoto, Com. III/I, pp. III.125-132. (B9) [22] Weisensee, M. [1988]: Model of Light Reflection for Facet stereo Vision. ISPRS 16th Congress, Kyoto, Com.III/4. pp. III.360-367. (B10) [23] SPOT. [1988]: SPOT User's Handbook. Vol.1 Reference Manual./ Vol.2 SPOT Handbook. [24] Wrobel, B.P. [1987]: Facet Stereo Vision (FAST Vision) - A New Approach to Computer Stereo Vision and to Digital photogrammetry. Conf. on Fast Processing of Photogrammetric Data, Interlaken, pp. 231-258. [25] Xiao, J.Z / Liu, J.Y. [1988]: Digital Matching of SPOT Stereo Images by Finite Elements Least Square Techniques. ISPRS 16th Congress, Kyoto, Com. III. pp. III. 373-382. (B10) [26] Zhang, Z .X. / Zhou, Y.Q. [1989] : A New Approach to Arrange the Approximate Epipolar Line for SPOT Images. Special Issue of Seminar on the Wang Zhizhuo' s Academic Thinking. Publishing House of Surveying and Mapping, Beijing. pp. 4650. 138